1019: LGTM, Ship It: The AI Code Review Problem

The hosts discuss the complexities and frustrations of working with code generated by AI, especially when trying to understand large and convoluted components. Tips include adding terse comments, establishing strict style rules, and smaller, more manageable prompts and changes. They emphasize the importance of standardizing style and avoiding reliance on effects and side effects in UI code. Additionally, the podcast highlights the increasing volume of PRs due to AI automation, which leads to technical debt, fragility, and architecture decay, stressing the ongoing importance of good software engineering practices despite the new automation tools.

The episode clarifies what local models are—AI models that run directly on user hardware versus those relying on API calls. They reference CJ’s detailed guide on local AI and discuss the significant compute requirements for running advanced models locally. The conversation explores emerging tools like jujutsu and DeltaDB, which aim to improve version control and conversation tracking in AI development workflows. These innovations seek to manage AI interactions more effectively and address the limitations of current systems, especially around storing context and managing revisions.

A discussion on how AI-driven development affects how contractors and freelancers should price their work. The hosts advocate for valuing the delivered outcome rather than hours spent since AI enables rapid feature creation. They stress that traditional effort-based pricing may no longer be fair or sufficient, and new models based on tokens or value delivered could be more appropriate. The episode underscores that despite technological advances, core principles of good software engineering—like managing technical debt and maintaining architectural integrity—remain crucial.

The hosts examine the WebKit box reflect property, noted for its visual appeal but lacking standardization and browser support. They argue that such trendy, CSS-only effects are often superficial and no longer relevant, especially with powerful alternatives like Canvas and WebGL. The discussion points out how CSS features tend to reflect obsolete trends and how developers should focus on higher-level, more flexible solutions rather than superficial visual effects.

The conversation compares the Angular framework’s reliance on built-in libraries versus the React community’s openness to external libraries. The hosts suggest that in the age of AI, leaning on browser standards and custom implementations may be advantageous, reducing dependencies and technical debt. They caution against the long-term maintenance costs associated with outdated or unused libraries, advocating for a balanced approach that leverages AI-generated code and standard APIs for more sustainable development.

Scott shares his experience building and fixing a Reachy Mini robot, emphasizing its expressiveness, programmability, and the DIY aspect of assembly. He talks about potential applications like education, learning assistance, or even gaming. Wes recommends his Bose QC35 headphones, which he has modified and maintained for over 10 years, highlighting their durability and quality. Both hosts showcase their personal projects and tech interests, ending the episode with enthusiasm about future innovations.